Architecture Without Architects, Again? Revisiting Vernacular Intelligence in the Age of Artificial Intelligence

Over sixty years after Rudofsky’s Architecture Without Architects, generative models offer architecture without architects of another kind, produced not by communities living with their buildings but by systems trained on their representations. This study asks how recorded vernacular knowledge is recognized, explained, and used in AI-mediated design. An inventory of fourteen principles of Cappadocian rock-cut settlements, each pairing form with cause, was fixed after a pilot, before the main runs. Claude Opus 5.5 and Claude Sonnet 5 completed a recognition, abstraction, and design task for Uçhisar under four conditions: place name, curated photographs, factual documentation without causal links, and pattern cards in Alexander’s format. Across 160 responses, coded by models without a human check, name-only responses showed a suggestive trend toward principles stated in public sources. Photographs made three use-related principles less frequent for Opus, whose knowledge recovery fell (0.62 to 0.48), while Sonnet shifted toward visible forms. Documentation increased causal recovery. Asked for sources, Opus said some cards exceeded the documentation; Sonnet favored carded principles. No literal imitation marker was detected in the written designs, but most principles also appeared in designs for the brief alone. The responses depended on how recorded knowledge was represented; lived processes remained unobserved.